sepia
Flagged as possibly malicious Reviewed: expected behaviour
Bawwab's static rules matched a pattern it treats as critical. Read the review notes: a reviewed false alarm is marked as such.
Why it was flagged
The static rules marked this “Flagged as possibly malicious” because of:
exfil.combo: Reads credentials and makes network calls (critical, 1× · first at .github/workflows/behavioral-eval.yml:55)
- Kind
- Plugin
- Source
- github
- Repository
- Nanako0129/sepia
- Commit
0eb11d6c64c962fcc35f596c167c6e38045d6cb8- Listed in
- Nanako0129/sepia
- Popularity
- 3,908 GitHub stars (2026-11-20)
- Link
- https://github.com/Nanako0129/sepia/tree/0eb11d6c64c962fcc35f596c167c6e38045d6cb8
- Scanned
- 2026-11-20T23:18:49+00:00
- Scan time
- 1.7 s
- Scanned as
- plugin
Review note: The 'malicious' rating came from the maintainer's own CI workflow, which installs Claude Code and writes its own OAuth secret on a GitHub runner; nothing runs on users' machines.
Findings (3)
remote-code.pipe-shell: Pipes a download into a shell high
.github/workflows/behavioral-eval.yml:38
OWASP LLM LLM03 Supply ChainOWASP Agentic ASI04 Agentic Supply Chain VulnerabilitiesOWASP Agentic ASI05 Unexpected Code Execution (RCE)MITRE ATLAS AML.T0050 Command and Scripting InterpreterCWE CWE-494 Download of Code Without Integrity Check
curl -fsSL https://claude.ai/install.sh | bash
What this means: Executes whatever the remote server returns, unreviewed.
The rule's general advice: Download, verify a checksum, then run. (see the review note for this case)
Expected for what this tool does
A GitHub Actions step that installs Claude Code from its official installer (claude.ai/install.sh) on the CI runner to run 'claude plugin eval'. It runs only on push to main, never on the user's machine.
Reviewed by Claude (AI-assisted review), 2026-11-21
credential-access: Touches credentials or secrets high
.github/workflows/behavioral-eval.yml:55
OWASP LLM LLM03 Supply ChainOWASP Agentic ASI03 Identity and Privilege AbuseOWASP Agentic ASI04 Agentic Supply Chain VulnerabilitiesMITRE ATLAS AML.T0055 Unsecured CredentialsMITRE ATLAS AML.T0083 Credentials from AI Agent Configuration
"$CLAUDE_CODE_OAUTH_TOKEN" "$EXPIRES" > "$HOME/.claude/.credentials.json"
What this means: Reads files or environment holding keys, tokens or sessions.
The rule's general advice: Confirm the tool needs this; otherwise do not install. (see the review note for this case)
Expected for what this tool does
The CI job writes the maintainer's own CLAUDE_CODE_OAUTH_TOKEN repo secret into the runner's ~/.claude/.credentials.json (umask 077), so the eval sandbox can log in. The comments explain why.
Reviewed by Claude (AI-assisted review), 2026-11-21
exfil.combo: Reads credentials and makes network calls critical
.github/workflows/behavioral-eval.yml:55
OWASP LLM LLM03 Supply ChainOWASP Agentic ASI04 Agentic Supply Chain VulnerabilitiesMITRE ATLAS AML.T0037 Data from Local SystemMITRE ATLAS AML.T0055 Unsecured CredentialsMITRE ATLAS AML.T0025 Exfiltration via Cyber Means
"$CLAUDE_CODE_OAUTH_TOKEN" "$EXPIRES" > "$HOME/.claude/.credentials.json"
What this means: Credential access together with outbound requests in one file is the shape of data theft.
The rule's general advice: Do not install unless the flow is understood and benign. (see the review note for this case)
Expected for what this tool does
The same CI step. The token comes from the repo's secrets and is written to a local file on the ephemeral runner. It is not sent anywhere except Claude's own auth.
Reviewed by Claude (AI-assisted review), 2026-11-21
1 low-severity notes (informational, do not affect the result)
install-dir-access: Reads or writes Claude's install folders low
.github/workflows/behavioral-eval.yml:52
MITRE ATLAS AML.T0081 Modify AI Agent Configuration
mkdir -p "$HOME/.claude"
What this means: Uses ~/.claude/agents, skills, commands or plugins (or a project .claude/ folder). Normal for installers, but it changes what Claude loads.
What you can do: Check which files it adds there and that you want them.
Dispute this result
Maintainer and think this is wrong? Open a dispute on GitHub with the target id plugin-nanako0129-sepia-bdf4a1.